Everything below was built and run on a single T4 GPU or local hardware. If a safety evaluation only works on a frontier-lab cluster, it isn't a usable guardrail.
Targeted RL experiments tracking how sparse feedback triggers erratic decision-making and reward hacking in autonomous agents.
Stress-tests 1.7B-parameter base models for reasoning blindspots — surfacing "right-answer, wrong-reason" behaviour that accuracy numbers hide.
RL as conditional sequence modeling — beats Nature DQN baselines by 20%+ from offline datasets, trained on constrained T4 compute.
A Claude-powered orchestrator that routes tasks to four specialist agents via tool use, in pure Python, with chained tool calls handled automatically.
High-throughput async evaluation framework (Groq, asyncio) that generates adversarial jailbreak scenarios, scores outputs across factual accuracy and safety compliance, and enforces a hard CI/CD gate that blocks deployment when safety scores fall below threshold.